AI Ethics & Operations

How Not To Build Skynet

TL;DR (The 30-Second Version)

AI is a tool, not a teammate. Humans stay accountable. Transparency beats magic. Data requires consent. Bias exists, plan for it. Document everything. Don't let models make decisions they can't explain. Review regularly. Unicorns are allowed; just keep them away from production.

What This Is (And Isn't)

This isn't a manifesto. It's not a shield against liability. It's the operating system for how I use AI at Sugartown and in my work, and if you're here reading this, probably how you should think about it too.

The technology is new. It's powerful, it's controversial, and the rules are still being written. That makes it more important, not less, to think about ethics and guardrails early, often, and in public. This page is a transparent accounting of where and how AI is used in my work, what I consider acceptable, and where I draw the lines.

Where AI shows up in this project

The accountability is mine. The code co-pilot is Claude. The thinking is a collaboration, but the signature on every deploy is a person's.

What this page isn't

The Operating Principles

1. Humans Stay Accountable — every decision traces back to a name, not a model version

AI can suggest, draft, summarize, remix, and hallucinate politely. Humans own decisions, outcomes, and apologies. If something ships, publishes, or impacts a person, a human signs for it.

In practice: Every commit in this repo carries a Co-Authored-By: Claude tag, but every merge to main is a human decision. Every Sanity document is published by a human. Node content is narrated by the AI agent, but Bex reviews, edits, and approves before it goes live.

2. Purpose Before Power — if grep handles it in 3 seconds, AI is cosplaying innovation

Use AI because it reduces toil, improves clarity, or unlocks creativity, not because it's shiny. If the AI output doesn't make work better, faster, or more humane, put the unicorn out to pasture.

The useful test: is this use of AI serving the work, or is it just optimising for dominance — speed, scale, competitive edge — as an end in itself? Those are different things with different consequences.

“It is not our part to master all the tides of the world, but to do what is in us for the succour of those years wherein we are set, uprooting the evil in the fields that we know, so that those who live after may have clean earth to till.” — JRR Tolkien via Magnifica Humanitas, Leo XIV, 2026

The argument isn't "do nothing grand." It's "do the right thing at the scale you actually operate." For an AI-assisted publishing platform, that means: ship with integrity, govern the system you've built, don't optimize for dominance. Clean earth to till.

(Personal note: this quote landing in a document about AI ethics is extremely on-brand for someone who grew up with two Galadriels and a Gandalf. Berkeley Girl Core™ activated.)

In practice: If you're using AI to solve a problem that grep handles in 3 seconds, you're cosplaying innovation.

3. Transparency Beats Magic — no deceptive automation cosplay

People should know when AI is involved, at least at a conceptual level. You don't need a blinking "AI DID THIS" banner, just no deceptive automation cosplay.

In practice: Knowledge graph nodes are explicitly AI-narrated (the agent is the "I" in the story). Articles are first-person PM, with AI collaboration credited. Every git commit is co-authored. The site's brand voice guide documents the split: articles = human narrator, nodes = AI narrator. No sock puppets.

Regulatory context (2026): The EU AI Act Article 50 transparency obligations become enforceable August 2026, requiring disclosure of AI interactions and labelling of synthetic content. Washington State signed AI disclosure and watermarking requirements into law in March 2026.

4. Data Is Not a Free Buffet — if you wouldn't paste it into a public doc, don't feed it to a model

Only use data you are allowed to use, understand the provenance of, and would feel okay explaining out loud to a lawyer, user, journalist, or your dad.

Consent matters. If someone didn't agree to have their data used for training, inference, or fine-tuning, you don't get to treat it like public domain just because it's technically accessible.

For what it's worth, the Pope agrees:

“ownership of data cannot be left solely in private hands.” — Magnifica Humanitas, Leo XIV, 2026

In practice: Client work stays in client contexts. Personal data never trains models. Scraped content gets side-eyed before use.

5. Bias Exists. Plan Accordingly. — assume bias is present, design review before harm

AI reflects the internet, history, and power structures, none of which are neutral. Assume bias is present. Design review, testing, and escalation paths before harm shows up in production.

In practice: Test outputs with diverse inputs. Challenge assumptions. If a model's answer feels off, trust your instincts and dig deeper.

6. Augment, Don't Replace Judgment — AI is a fast intern with perfect recall and zero lived experience

AI is great at pattern detection, drafting, summarizing, and exploring possibilities. AI is bad at moral reasoning, contextual nuance, accountability, and formatting output according to spec.

Treat it like a very fast intern with perfect recall, zero lived experience, and a tendency to confuse "sounds right" with "is right."

In practice: AI can draft the spec. Humans decide if it ships. AI can suggest copy. Humans verify it's not confidently wrong.

7. Fail Softly — no silent automation cliffs, no "oops, the model decided" moments

Design systems so AI mistakes are reversible, inspectable, and non-catastrophic.

In practice: Sanity content is created as drafts; a human publishes. Migration scripts run in dry-run mode first (always). Validators (validate:tokens, validate:urls, validate:content) catch drift before it ships. Schema changes require explicit deployment.

8. Governance Is a Feature — if you can't explain why the AI did something, you don't understand your system

Logs, versioning, audits, prompts-as-artifacts, and clear ownership are not bureaucracy. They are how you prove you're not running a chaos engine with vibes. Prompts are documentation. Versioning is a safety rail.

On external accountability: Governance isn't only internal. If your AI system affects users, they should have a meaningful way to flag issues and influence how the system behaves — not just a contact form that goes into the void (I have one of those, here). Communities affected by AI decisions shouldn't be passive recipients of choices made elsewhere.

In practice: CLAUDE.md defines conventions the AI agent must follow. MEMORY.md persists architectural decisions across sessions. Epic specs in docs/backlog/ are the prompts that drive implementation. Post-mortem findings are codified as rules, not just learned.

Industry context (2026): ISO/IEC 42001 (AI Management System) is the auditable standard for AI governance. Regulators are asking for demonstrable controls, not aspirational ethics statements.

9. Creative Play Is Allowed — unicorns are welcome, just keep them away from production

Yes, you may brainstorm with AI, use metaphors, generate unicorns, diagrams, thought experiments, and hilarious self-portraits. Just don't confuse play with truth or fiction with policy.

In practice: AI-generated mockups, diagrams, and concept art? Great. AI-generated legal advice, financial projections, or hiring decisions? Absolutely not.

10. Revisit Regularly — responsible AI is not a checkbox

Models change. Context changes. Stakes change. Review assumptions like you review dependencies.

In practice: This page gets reviewed every 6 months. Principles update when failures teach us something. The changelog is public.

11. Attribution Matters — credit your tools like you'd credit a co-author

If AI helps write, design, or code something, that's collaboration, not plagiarism, not magic. If a model trained on scraped data produces something suspiciously specific, verify provenance.

In practice: Every git commit carries Co-Authored-By: Claude. The brand voice guide documents which content is AI-narrated versus human-narrated. The site itself is the portfolio for this collaboration, not a secret.

12. Compute Has a Carbon Cost — efficiency is an ethics problem, not just an ops problem

Every API call burns electricity. Every training run has a footprint. Use AI where it creates value, not where it's a lazy substitute for a search query or a bash script.

In practice: Use cached results when possible. Don't regenerate content unnecessarily. Batch API calls. Choose smaller models when they suffice.

13. AI-Generated Content Requires Labelling — if a machine made it, say so

As AI-generated text, images, and media become indistinguishable from human-created content, disclosure is a baseline ethical obligation, not just a legal one.

In practice: Published articles drafted with AI include a disclosure. An image flagged as AI-generated gets a visible "AI-generated" label on the page. It is also served as a copy with the label baked into the pixels and metadata embedded in the file (IPTC digital source type, plus the tool and version when known). The social share image gets the same treatment. The node system credits the AI agent as narrator by design; articles credit AI collaboration in the byline.

What this does not do: anyone can crop or re-save a file and lose the label and metadata. We do not sign images with C2PA Content Credentials or add an invisible watermark, because both need a signing product that has passed C2PA conformance, a certificate from a trusted authority, and a lookup service, and we run none of them. If an image arrives with its generator's own Content Credentials, our re-encoding drops them.

Regulatory context: EU AI Act Article 50 applies from 2 August 2026. Providers of generative AI systems must mark their outputs in a machine-readable way. Deployers must disclose deepfakes and AI-generated text published to inform the public, with lighter disclosure for evidently artistic or creative work. Our labels and metadata go beyond that and are voluntary; this page makes no claim of compliance. US states are enacting similar disclosure laws.

When Things Go Wrong (Not If)

AI will fail. Models will hallucinate. Bias will surface. When that happens:

  1. Stop the system. Pull it offline if harm is occurring.
  2. Document the failure. Logs, inputs, outputs, context.
  3. Communicate transparently. Tell affected people what happened.
  4. Fix the root cause. Not just the symptom.
  5. Update this page. Your ethics are only as good as your learning loop.

Escalation path: Issues go to contact bex. Critical failures pause all automated systems until root cause analysis completes.

Confirmed failures are recorded in a standing incident log, and the project's full AI-governance posture (what's owned in code, what's inherited from platforms, and what's out of scope by design) is mapped on the Platform Governance page at /platform/governance.

Accessibility Commitment

AI Must Not Break Accessibility. If AI generates content, it must:

If a model outputs inaccessible garbage, that's a bug, not a feature.

Canonical References (When You Want Receipts)

i1: Standards & Frameworks (NIST, OECD, ISO)

NIST AI Risk Management Framework (nist.gov): Practical, risk-based, and refreshingly grounded.

OECD AI Principles (oecd.ai): The global baseline most policies quietly borrow from.

ISO/IEC 42001 (iso.org): The auditable management system standard for AI governance.

G7 Hiroshima AI Process: International consensus on AI governance principles.

i2: Regulatory & Policy (EU AI Act, US & state regulation)

EU AI Act (ec.europa.eu): Prohibited systems banned Feb 2025. GPAI transparency Aug 2025. Article 50 content labelling Aug 2026.

US Federal & State Regulation: The regulatory landscape is actively shifting. Colorado AI Act, Washington AI disclosure/watermarking (March 2026), California AI transparency bills. Cite a specific law only when it directly governs your operations.

i3: Humanitarian & Institutional Ethics (Magnifica Humanitas, Rome Call)

Pope Leo XIV — Magnifica Humanitas (2026): The Catholic Church’s first full encyclical on AI. Secular or not, the governance argument is serious: data ownership, participatory oversight, AI in warfare, and concentration of decision-making power. (vatican.va)

Vatican — Rome Call for AI Ethics (2020): Signed by Microsoft, IBM, and Cisco. Established the Vatican’s position on human dignity and algorithmic accountability before it was fashionable. (romecall.org)

i4: Model Providers & Safety Research (Anthropic, OpenAI)

Anthropic Responsible Scaling Policy (anthropic.com): How the people who built Claude think about catastrophic risk. Since we use Claude extensively, worth knowing their red lines.

OpenAI Safety (openai.com): Useful for understanding how frontier-model builders think about risk.

i5: Practical Case Studies (Partnership on AI)

Partnership on AI (partnershiponai.org): Real-world examples of AI doing weird, unexpected things in production.

i6: Licensing & Copyright (Creative Commons)

Creative Commons AI & Licensing (creativecommons.org): What copyright means when machines do the remixing.

The Quiet Truth Beneath the Cheeky Attitude

Responsible AI isn't about neutering creativity or slowing teams down. It's about making AI boring in the ways that matter, so it can be powerful where it helps.

You can have unicorns. You can brainstorm with models, generate absurd diagrams, and let Claude write your first draft.

Just don't let them run payroll, justice, hiring decisions, or prod deploys unsupervised.

And if something breaks? A human apologizes. Not a model.

Changelog

v2026.10.31 — Share Images and an Article 50 Correction

Social share images can now carry the same label and embedded metadata as page images, and stamped files record the tool version when known. The tool list now covers ChatGPT, Claude, Claude Code, Gemini, Adobe Firefly, Photoshop Generative Fill, Canva AI, Microsoft Copilot and Stable Diffusion. Corrected Principle 13: the EU marking duty sits with providers of generative AI systems, not with publishers. Known gap: stamping re-encodes images, which drops any Content Credentials the generator embedded.

v2026.10.30 — AI Image Labelling

AI-generated images can now carry a visible "AI-generated" label on article images, hero images, card thumbnails and gallery images, with the tool named on hover or focus. Flagged images are also served as copies stamped with embedded metadata (IPTC digital source type and AI system used). Not done: C2PA signing and an invisible watermark. They need a signing product that has passed C2PA conformance, a certificate from a trusted authority, and a lookup service, none of which we run.

v2026.06.01 — June Review (Magnifica Humanitas Update)

Added Humanitarian & Institutional Ethics sub-section to Canonical References, including Magnifica Humanitas (Leo XIV, 2026) and the Rome Call for AI Ethics (2020). Strengthened Principle 4 with data ownership note drawn from encyclical. Added external accountability note to Principle 8 on participatory governance. Expanded Principle 2 with Gandalf quote (§213, Magnifica Humanitas), contextual gloss, and personal aside. Restored Principle 13 as standalone principle with distinguishing note. Added US Federal & State Regulation catch-all to Regulatory & Policy references. Bumped next review to December 2026.

v2026.04.11 — Structural Revision

New principle #13 (AI-Generated Content Requires Labelling). Regulatory context added to #3 (Transparency) and #8 (Governance). Updated canonical references. Opening statement rewritten with transparency rationale. Page restructured with accordion sections. Next review: October 2026.

v2025.12.29 — Initial Public Version

Established 12 core operating principles. Added incident response protocol. Defined accessibility commitments. Compiled canonical reference library. Set 6-month review cadence. Formalised existing practices into documented policy.

Last updated: October 31, 2026. Next review: December 2026. Questions? Contact Bex (Human. Probably.)